Rainbrand Skill and Agent Plugin

Rainbrand is the brand layer for agents. It turns a brand's identity, guidance, and assets into shared context that agents and products can use and update as the brand evolves, alongside the tools where it is created and managed.

This repository maintains one optional Agent Skill and packages it with Rainbrand's remote Model Context Protocol (MCP) connection as an Agent Plugins 1.0 plugin. Both use the identifier rainbrand.

The Rainbrand Skill teaches an agent when and how to use Rainbrand. A Brand Skill contains the context for one particular brand. Installing the standalone Skill supplies guidance; the plugin also declares the MCP connection. The client handles authentication.

Install

From this checkout, install the Skill in the current project:

npx skills add . --skill rainbrand

Prefer a project installation so the guidance is reviewable with the project.

Standalone Skill upload

The local dist/rainbrand.skill.zip contains the canonical Skill as rainbrand/SKILL.md. Upload that archive through Claude's Skill upload flow.

Portable plugin

Use this repository root as the plugin directory, or extract the local dist/rainbrand.plugin.zip. Load it in a client that supports Agent Plugins 1.0 and Streamable HTTP, then follow the client's OAuth sign-in flow.

The archive contains these files directly at its root:

plugin.json
mcp.json
assets/rainbrand-logo.png
skills/rainbrand/SKILL.md
README.md
CHANGELOG.md
LICENSE

Both manifests declare the Agent Plugins 1.0.0 schemas. mcp.json defines one server named rainbrand with transport streamable-http and endpoint https://mcp.rainbrand.com/mcp. The package contains no credentials. assets/rainbrand-logo.png is the approved temporary white Rainbrand R on black, used for both the OpenAI listing logo and composer icon.

Connect Rainbrand

For interactive work, connect https://mcp.rainbrand.com/mcp in an MCP client that supports Streamable HTTP and OAuth. A plugin client reads that connection from mcp.json; a standalone Skill installation needs it configured separately. Complete Rainbrand sign-in, then let the client discover the live tools and their schemas. The live tool list determines what the connected account can use.

For product information, support, and API integration guidance, use the Rainbrand website and FAQ. Before building an application or backend integration, obtain a current, verified Rainbrand API contract. Keep API keys in the application's secret store.

Provider compatibility

The portable package is a draft for client testing. Successful local validation does not establish acceptance by a provider or publication in its directory.

OpenAI's build and submission requirements are documented in Build plugins and Submit your plugin. Its review rules require a new submission when the MCP server origin changes; this migration must not update the previous submission's origin. Test the actual target client; importing this ZIP does not establish ChatGPT web availability.

Claude's native format requires .claude-plugin/plugin.json and .mcp.json with transport http. Those root paths conflict with the portable specification's requirement that client-specific files use a reverse-domain top-level directory. This repository does not add an undocumented adapter or a second maintained implementation. Reconcile those contracts before adding a native adapter. At that point, also check Claude's archive rules: the repository contains distribution ZIPs, and a native upload must not contain nested ZIPs. See Claude plugin requirements, the pre-submission checklist, and directory publishing.

Cursor recognizes the root plugin.json; a separate .cursor-plugin/plugin.json manifest is not required for this format. See Cursor plugins and marketplace submission.

Review and publication checklist

The manifest supplies five read-only positive cases and three negative cases for OpenAI review. Tool names match the server's raw tools/list names: list_brands, get_brand, get_account, list_workspaces, and check_credits. A client may qualify them with the server name, for example mcp__rainbrand__list_brands.

Prepare a dedicated reviewer account with a sample Stripe brand. Run the positive cases in order in the same conversation: the first case supplies the Stripe Brand session ID for get_brand, and the workspace case supplies the personal workspace_id for check_credits. The personal-credit case expects only check_credits. The exchange-rate, Tokyo weather, and arithmetic cases must invoke no Rainbrand tools; the arithmetic answer is 42. Live connection, discovery, and calls to all five tools have passed. The walkthrough exercises all eight exact prompts with the dedicated reviewer account and Stripe brand. Put reviewer credentials and private instructions in the provider's review portal, never in this repository.

Before public submission:

• The manifest includes demo_recording_url pointing to the Rainbrand walkthrough, published with the release.
• If OpenAI issues a new ownership challenge, serve the exact plain-text token from the portal at the eligible Rainbrand HTTPS origin. Do not reuse an old challenge token.
• Check public website, support, and legal pages for the completed Rainbrand rollout. Reachable pages alone do not establish that launch is complete.
• Complete each provider's client tests, review requirements, and submission process. Provider review determines whether these brand and design workflows are accepted; the portable format is not universal approval.

A later repository rename is optional housekeeping, not a submission gate. Release v1.1.2 includes both distribution ZIPs and the review walkthrough. Preserve every earlier tag, release, and asset unchanged. Use the current public repository URL in provider portals that require one.

Update or remove

npx skills update rainbrand
npx skills remove rainbrand

Add --global for a global installation.

Package and validate

The canonical source is skills/rainbrand/SKILL.md. Build both archives and validate them in this order:

scripts/package-skill.sh
scripts/package-plugin.sh
scripts/validate-plugin.sh

Pass a ZIP path to scripts/validate-plugin.sh to validate another local copy against this checkout. Check submission readiness separately:

scripts/validate-plugin.sh --submission

The draft package can pass ordinary validation while submission readiness fails for the missing approved artwork and recording metadata. Readiness validation does not replace the manual publication and provider checks above.

The scripts run offline with Bash 3.2 or later, POSIX awk, zip, Info-ZIP unzip (including -Z), and standard shell utilities. They require no package manager or additional language runtime.

Validation checks JSON types, duplicate and unsupported fields, schema identifiers, matching names, the exact MCP endpoint, and OpenAI metadata. The plugin version, Skill version, and latest numbered changelog release must agree. Archive checks cover integrity, exact members, regular file types, and byte-for-byte agreement with source files. Both builders fix timestamps and file modes so unchanged sources produce identical archives with the same ZIP tool version, including across time zones and umasks.

For packaging changes, run the optional regression suite:

python3 scripts/test-validation.py

These tests exercise malformed metadata, unsafe or mismatched archives, submission gates, and reproducible builds. Python 3 is required only for this test suite, not for package building or validation.

License

MIT